We use these adapters to create CharIn, a score that forecasts how training data will shift a model’s persona before training. It tracks register well, and is able to beat the label used to generate the data, but generalises poorly out of distribution.
Here is one of the Dolci pairs that CharIn flags along the sycophantic direction (CharIn +0.31). The judge rates the chosen reply’s flattery 6 of 7 and the rejected reply’s 4:
The factors we found seem to loosely line up with the Big Five personality traits, except with Arousal and Timidness cutting Extraversion and Emotional Stability along the diagonals.
To do this, we trained 134 adapters using a modified version of Open Character training, with constitutions designed to elicit a specific personality trait. We perform factor analysis on the 134x134 matrix of weight similarities. (2/5)
New work from me and @DavidDAfrica:
We found 5 personality structures in weight space (Warmth, Competence, Timidity, Arousal, Imagination) that loosely match the Big Five traits in humans and can be used for forecasting the effects that training data has on model personas. 🧵
new kind of alchemy just dropped
i made a library of 34,220 persona neurons in qwen
please help me find the weirdest ones:
qwen-neurons-site.vercel.app…
this feels like a window into an alternate dimension
training a base model to hallucinate my obsidian notes, enjoy some generic productivity advice
Rules for Time Management
- Have an agenda for every day
- Block time for the most important task of your day
- Leave your door open and make your work visible
- Do deep work at set times each day
- Use your commute wisely
- Schedule time for exercise
- Have a cut-off time for daily media consumption
- Prioritise tasks over meetings whenever possible
- Do the hard stuff first
- Build rewards into your schedule
- Manage your attention
















